What is an adversarial example in the context of AI?

• A. Input that improves model performance
• B. Specially crafted input causing incorrect output
• C. Data used for training and validation
• D. Random noise added to inputs

Answer: (B)

An adversarial example refers to a specially crafted input designed to deceive an AI model, leading to incorrect output. These inputs are manipulated in specific ways, often imperceptible to human observers, but enough to mislead the machine learning algorithms into making errors in classification or prediction.

This phenomenon is significant in AI, particularly in domains like image recognition and natural language processing, where subtle changes can confuse the model. For instance, in image classification, altering a few pixels in an image can cause a model to misidentify it entirely, which illustrates the vulnerability of AI systems to adversarial attacks. Understanding adversarial examples is crucial for developing robust models that can withstand attempts at manipulation. This is why this choice is accurate and relevant in the study of AI security and resilience.